A language model can always produce a reason. That doesn’t make the reason the cause. Anthropic tested this directly in 2025: it slipped hints into prompts that changed models’ answers, then checked whether the models admitted using them. Claude 3.7 Sonnet mentioned the hint about 25% of the time. DeepSeek R1 managed about 39%. The rest of the time, the models wrote tidy justifications that left out what actually swayed them.

Repeating the question a hundred times doesn’t fix this. It tells you which excuses a model likes to give, not what drives its choices. If the explanation is a rationalisation, a hundred of them is just a well-sampled rationalisation.

The Anthropic work tested reasoning in controlled question-and-answer tasks, not AI search. But the problem goes deeper than honesty. Even a perfectly faithful explanation couldn’t report what the model never saw: retrieval scores, ranking decisions, system instructions, or the pages filtered out before the answer was written. Those happen in separate systems. A generated explanation is not an audit log.

A 2026 study gave eight models 2,000 product queries and shuffled the order of the candidate list 20 times. Every model’s picks depended on that order: the top recommendation changed in 15.5% to 34% of cases from reshuffling alone. Ask a model why it chose, and you’ll get a reason about the product, not “it was listed first”. That’s exactly the kind of influence the Anthropic work shows models rarely own up to.

That number measures one thing: how often the model said yes to that exact question, worded that way, at those settings. It is not the chance the model recommends the brand to a real customer.

• You named the brand. A real user asking for an accountant never does. You’re testing agreement, not recall.
• The wording drives it. Rephrase the question and the rate moves.
• The settings drive it. Temperature and sampling change how often you get yes. That’s the dial, not the brand.
• Nobody has calibrated it. There’s no evidence this rate predicts what happens in a real, search-grounded answer.

Call it what it is: a yes rate for a leading question. Renaming it “confidence” makes it sound like you’ve measured the model’s opinion of a business. You haven’t.

A CIKM 2026 paper found synthetic prompt sets overstated retrieval performance by 29 to 37 points against real user queries. And a 2026 methods paper showed that re-weighting prompt groups in published data moved an aggregate rate, in this case how often an AI answer appeared at all, from 39.7% to 70.5% with nothing inside any group changing. It isn’t a brand visibility figure, but the lesson carries straight over: which prompts you sample, how you weight them and how the model is configured decide what your probe measures.

Order of mention is not order of preference. Read this answer:

“Brand A is cheaper and Brand B is well established, but Brand C is the best option if reliability matters.”

Brand A is mentioned first. Brand C is the recommendation. A tracker that scores by mention order puts the loser at the top.

Or this one: “Brand A and Brand B are both excellent choices.” There’s an order of names. There’s no order of preference.

Position does mean something when the answer explicitly ranks its picks: “here are my top five, in order”. The myth is treating every sequence of brand names as a ranking.

Mention and recommendation are different outcomes. And the positional-bias study in Myth 1 shows the order brands arrive in can shift which one gets picked, so even the order of names can be an artefact of the input, not a judgement.

What was shown is that one model, acting as a judge in one test, ranked one snippet higher without that line, against one set of competitors. That’s an observation about a test. It is not a ranking factor in Google AI Mode, ChatGPT or anywhere a customer actually searches.

The result could come from the competitor set, how dense the snippet became, the other brands’ ratings, where the text sat, or plain run-to-run noise. A judge model is also not the production system. Optimise against your own judge for long enough and you get very good at pleasing your own judge.

The large-scale data points elsewhere on star ratings.

These are associations, not causes. Brands with more reviews also tend to have more customers, more search demand and more coverage, and neither study isolates review count as the lever. But neither supports a star rating as a negative ranking factor either.

The favourite formatting tips fare no better. Chen and Koudas studied 468,037 citations and found FAQ blocks, schema, author boxes, summaries and page speed showed no reliable advantage. What did count was relevance: paragraphs and titles that closely matched the query were linked to 14% to 16% more citations. A feature that helps a page match the search is a different thing from a feature the model rewards for its own sake.

The tell is in the workflow itself. These tools hand you a content brief, and you then have to change the live site and wait to see if anything moves. That’s a hypothesis generator. Fine to use, but don’t sell its outputs as discovered ranking factors. Turning a hypothesis into a finding takes controlled changes on live pages, repeated runs and checks against the AI systems customers actually use.

The same slip shows up in a related claim: if you’re the second grounding source, you’re second in organic search. Not necessarily. AI answers often fan out into several searches, and the order sources arrive in the model’s context isn’t the same thing as a search position for the user’s query.

Ask a model to invent reviews for a little-known dental practice and you’ll get long waits, rushed appointments and surprise fees. That’s not your brand. That’s what dentist reviews look like on the internet. Proponents usually concede that for small brands the output reflects the industry, not the business. Then they argue the invented negatives could hurt you anyway.

For an unfamiliar brand, the invented reviews are mostly industry pattern. For a well-known one, they may mix in things the model genuinely learned about the brand. The trouble is you can’t tell which is which. Without comparing the output against real reviews, competitors and actual recommendation outcomes, you can’t say how much of it is about you, or whether any of it predicts being recommended. I haven’t seen anyone publish that test.

This isn’t direct evidence against the method, but it’s a reminder of where AI systems actually get their evidence about you: real, third-party content.

What this measures is which brand names one model tends to produce under one prompt. Coca-Cola turning up every time tells you Coca-Cola is famous in English-language text. You already knew that.

The expansion step bakes in bias. Start from the most famous Western brands, ask for “related” brands, and the map you get is shaped by where you started. If brands from one country or sector rarely appear, that’s a property of the seeds, the prompt and the training mix. It isn’t evidence the model “dislikes” anything. Models don’t have grudges; they have distributions.

Malthouse and colleagues found many established brands received no category-level recommendations at all, despite models clearly knowing them. Being known and being recommended are separate.

There’s a real point buried here. Repeated brand prompts can usefully show which associations a model reaches for, and fame clearly shows up in models. The mistake is calling that output an authority score. You build fame in the world, then check whether it’s showing up in answers to real buyer questions.

The labs say otherwise. Anthropic’s privacy policy lists publicly available internet data first among its training sources, alongside licensed datasets and user data. And since September 2025, Anthropic has asked consumer users to choose whether their chats can be used for training, with the setting on by default. Curated and synthetic data have grown. The web hasn’t left the building.

There’s a fair point inside the myth: a single chat doesn’t retrain the live model, and nobody can game the global model by clicking or prompting at it. The leap to “training data no longer matters” isn’t.

An uncited mention isn’t proof on its own that a brand came from training data. It could have come from a page the model read but didn’t cite. But the rest of Chen and Koudas’s analysis points the same way, and so does other work.

A pre-registered Gemini study found memory-based brand mentions were still predictable 20 days later. Another found that facts retrieved, and even cited, could still lose to stale model memory. Training data isn’t irrelevant. It’s the incumbent.

It’s also self-contradictory. The same people usually say models carry a learned bias towards famous brands that shapes recommendations. That bias came from somewhere. It came from training data, and the web is still in it.

There’s no source for 90%. No dataset, no platform mix, no method. It’s a feeling dressed as a statistic.

Retrieval clearly matters for many queries, especially local and commercial ones. But “unless you’re a major brand” is doing a lot of work in that sentence. It concedes the whole point: being known changes what the model does. The same camp argues the model re-ranks what search hands it using its own judgement, that brand familiarity tips that judgement, and that a poor reputation in the model can drag you down even when you rank. Those are all effects that sit outside search rankings.

A claim like that needs a definition of visibility, a platform mix and a method you can repeat. Without them, it can’t be tested. The strongest research we have, from Chen and Koudas, finds retrieval central to which pages get cited, and also finds model knowledge shaping which brands get searched for and named. That describes visibility built in several stages, with search and model knowledge feeding each other. It doesn’t support a single number that hands search nearly all the credit.

Being findable is the entry ticket, and it isn’t hard to earn. What it doesn’t do is get you chosen.

An AI answer isn’t a list of ten blue links. It’s a judgement about which brands the evidence supports, and most of that evidence lives on other people’s sites.